Defining groups in longitudinal data + other assorted questions

[chiuhoward]

Hello pNet team, I had a few questions about the hierarchical volumetric pNet.

• are the group-level statistics presented in the html outputs stored anywhere as data files? i've looked through the output directory and could only find individual-level outputs.

• for the group-level representations, is the group defined as simply whoever i've included in the list of files for pNet? for context, i am analyzing a longitudinal reading intervention dataset, and i grouped only by session so the participants in this folder includes both children with and without reading difficulties. there are data for multiple sessions, but i ran pnet on each session separately by providing the different session folders to separate instances of the pNet container. however, the number of subjects in each session would be different as i may exclude some sessions from different subjects due to motion etc.

• for quality assurance, each subject gets a QA txt output, but this is merely descriptive? i.e. pNet does not exclude any runs based on the QA/QC metrics because it assumes fmriprep or whatever preprocessing came before would have done all the quality control steps? beyond the 6mm FWHM smoothing, is there a recommended cleaning pipeline for motion confounds etc?

• in the paper, 'HCsall’ is defined as combinations of ‘HCs148→50’ and ‘HCs50→17’, is that just a literal concatenation of both sets of features or there was some other transformation involved?

Thank you!

[MLDataAnalytics]

Please see our responses below:

The group-level FNs are included in the Docker image (in the model folder) and are used as surrogates for computing the QA metrics. If your data differ substantially from the HCP young-adult cohort, these QA results may be less informative.

The group-level representations are simply the average of the FNs across all individuals in the training set. The tool generates individualized FNs and does not explicitly rely on the group-level representations during inference.

The QA results are quantitative and include functional heterogeneity measures and spatial correlations. Functional heterogeneity is computed separately for each FN for both individualized and group-level FNs, and the resulting values are then compared. Spatial correlation is computed between individualized and group-level FNs and compared across FNs.

HCsall is the concatenation of the two feature sets. In addition, a new feature set can be generated via matrix multiplication between the two sets (e.g., HCs148→50 × HCs50→17).

Thank you for these questions. They are very helpful for improving the pNet documentation.

[chiuhoward]

Thank you very much for the previous clarifications! They were very helpful. To ensure I am implementing the hierarchical volumetric pNet correctly, I have a few specific follow-up questions:

1. Group-Level Statistics Export
   In the HTML output, pNet presents summary statistics and visualizations for my specific cohort. While I understand the HCP templates are stored in the Docker image, are the statistics generated from my data (e.g., the group-level means/spatial maps shown in the report) saved anywhere as a .csv, .mat, or .nii.gz file? Or should I plan to calculate these group-level representations manually from the individualized outputs?

2. QA Metric Logic: Descriptive vs. Analytical
   Regarding the QA results (functional heterogeneity and spatial correlations): Does pNet automatically exclude subjects or runs from the pipeline if they fall below a certain threshold? Or is this output purely descriptive, leaving it to the user to perform manual exclusion post-hoc based on the .txt report?

3. Recommended Denoising & Preprocessing
   Since pNet assumes prior preprocessing, is there a "gold standard" denoising pipeline you recommend that pairs best with the pNet architecture? Specifically, beyond the 6mm FWHM smoothing, do you recommend a particular approach for:

• Nuisance regression (e.g., 24-parameter motion, CompCor, or ICA-AROMA)?
• High-pass filtering (e.g., > 0.01 Hz)?
• Global signal regression (GSR)?
  I am leaning towards using fMRIprep —> XCP-D —> pNet to address this?

4. Longitudinal Consistency & Varying Scan Lengths
   My study has an unbalanced longitudinal design with some subjects having more sessions than others. Since I am running separate instances of the pNet container for Session 1, Session 2, etc., will the individualized FN estimations remain consistent/comparable for longitudinal analysis?

Thank you again for your time and for this excellent tool!

[MLDataAnalytics]

1. Group‑Level Statistics Export

pNet currently does not automatically export group maps though the summary maps and statistics are shown in the HTML report.

The individualized maps (per subject/session) are saved in standard formats (.nii.gz for spatial maps), and you can compute group maps from the individualized outputs. These are directly compatible and already in the same space, so you can average voxel‑wise across subjects.

We will add an option in a future release (e.g., --export-group-stats) to automatically produce: group mean / SD maps for each FN (.nii.gz)

2. QA Metrics: Descriptive vs. Analytical

The QA outputs (functional heterogeneity, spatial correlation with templates, motion summaries, etc.) are post-hoc and users can inspect the QA results and apply their own study‑specific criteria to remove low‑quality subjects if needed.

3. Recommended Preprocessing & Denoising Workflow

pNet assumes the functional data are already preprocessed and denoised. Several pipelines work well, but we generally recommend: fMRIPrep → XCP‑D → pNet

Spatial smoothing: 6 mm FWHM (default). 4–6 mm is generally appropriate.

Nuisance regression: 24‑parameter motion model + aCompCor (5–10 components), or ICA‑AROMA (non‑aggressive) + aCompCor or motion derivatives

High‑pass filtering: ≥ 0.008–0.01 Hz, Low‑pass (optional): ≤ 0.1–0.15 Hz if desired, Global signal regression (GSR): optional and study‑dependent

4. Longitudinal Consistency with Varying Scan Lengths

pNet model was trained to be robust to varying scan lengths. To make the longitudinal analysis not biased to scans with different lengths, it is recommended to apply pNet to scans with the same length. You can also run a comparison between pNet results from the same fMRI scans but with different lengths.